mHC coding scheme boosts semantic communication with zero extra bandwidth

Semantic communication (SemCom) and task-oriented communication (TOC) aim to slash wireless resource consumption by transmitting only the meaning or task-relevant bits, rather than raw data. But existing learning-based transceivers often compensate for channel noise by scaling up encoder dimensions or feature size — which increases computational complexity and channel usage. In a new paper on arXiv (2608.13253), Jingwen Fu and Ming Xiao from the Department of Computer Science and Information Theory propose a different approach: a manifold-constrained hyper-connection (mHC) coding scheme that makes representations both compact and resilient without extra bandwidth.

The mHC encoder replaces the standard single residual path with multiple residual streams, mixing them through doubly stochastic (DS) matrices. This constrained interaction boosts representation diversity and training stability, with negligible parameter and floating-point overhead. An entropy bottleneck (EB) then quantizes the transmitted channel features and estimates an entropy-coded rate, enabling end-to-end rate–distortion/task optimization under real bandwidth and power constraints. The authors prove that DS-constrained mixing does not increase the differential entropy of the transmitted features — meaning it doesn't lengthen the ideal entropy coding either. In experiments across additive white Gaussian noise (AWGN), Rayleigh fading, Rician fading, and imperfect channel state information (CSI), mHC outperformed both residual and unconstrained hyper-connection baselines in semantic/task performance, robustness, and convergence, all while using no additional channel uses.